QA for Natural Disaster Scenarios: An Efficient Approach

Emergency management in the event of natural disasters requires rapid and accurate responses. A new study presents a question answering (QA) system specifically for disaster scenarios, focused on typical situations and responses in Japan.

The approach is based on the cl-tohoku/bert-base-japanese-v3 architecture enhanced with Bi-LSTM and optimized using LoRA (Low-Rank Adaptation). This technique drastically reduces the number of parameters required, achieving high computational efficiency.

Implementation Details and Performance

The model, with only 5.7% of the total parameters (6.7M out of 117M), achieves 70.4% accuracy in identifying the end position of the answer. Experimental results demonstrate that the combination of Japanese BERT-base and Bi-LSTM for contextual understanding achieves accuracy levels suitable for real disaster response scenarios, with an F1 score of 0.885.

Future Perspectives

Future steps include creating benchmark datasets for QA in the field of natural disasters, fine-tuning foundation models with domain-specific knowledge, developing lightweight and power-efficient edge AI applications for situations with limited power and communication, and implementing continuous updates to the knowledge base.